Proactive Bandwidth Orchestration in Industrial SD-WANs: A Time-Aware Deep Reinforcement Learning Approach

Abstract

The aggregation of concurrent Industrial Internet of Things (IIoT) data traffic from distributed sites to the cloud frequently induces uplink congestion at the capacityconstrained cloud infrastructure ingress. Conventional reactive congestion control mechanisms are constrained by inherent feedback delays, while static bandwidth allocation strategies fail to adapt to stochastic traffic fluctuations, resulting in resource underutilization during off-peak periods and packet loss during bursts. To address these limitations, we propose a recurrent deep reinforcement learning (DRL) framework for proactive bandwidth orchestration. By integrating gated recurrent units (GRUs) into the policy architecture, the framework extracts latent temporal features from historical network observations, thereby capturing the underlying dynamics of bursty traffic. Driven by these temporal features, the DRL agent proactively regulates gateway transmission rates, mitigating ingress congestion and protecting mission-critical flows against packet loss. Performance evaluations under realistic industrial traffic profiles demonstrate that the proposed framework reduces sample complexity by over 50% compared to memoryless DRL and heuristic baselines, while constraining the ingress packet loss rate to $\mathbf{1. 2 0 \%}$ under heavy saturation.